Problem

Training a computer vision model to detect objects in operating rooms is complex, due to the data being protected health information (PHI). HIPAA regulations require strong data protection measures at every stage of the computer vision pipeline. Large datasets have to be encrypted and curated for annotation, and a computer vision model viable for assisting in live surgery has to be highly accurate.

Solution

LayerNext’s MetaLake hosts the data protection and management process from encryption to curation, annotation, and model training on one unified platform. Its cloud-based solution runs internally on a customer’s existing cloud infrastructure to ensure data stays encrypted in its servers, yet is accessible through the computer vision pipeline for model training.

Result

With compliance regulations and data operations largely managed by the LayerNext platform, the company had increased capabilities to focus on model training and accuracy. Leaning on LayerNext’s model performance benchmarking, the data engineering team was able to develop the AI model at 95% model accuracy, and at 4x the efficiency rate.

A technology company explores the use of computer vision to detect objects in Operating Rooms to improve patient safety. Working with image and video formats, they needed large datasets that could capture the dynamics and complexity of live surgery. To fully realize an AI model that can be deployed in different hospitals and operating theaters, they needed a software platform that can offer data protection in all stages of the computer vision pipeline, remote data management, and features that improve AI performance. Here’s how LayerNext met their needs.

Data protection and HIPAA compliance

The technology company needed to work with protected health information (PHI), while having strong data protection measures in place, and following them to ensure HIPAA compliance. With LayerNext’s cloud-based platform, they are able to run the MetaLake internally on their customer’s infrastructure so that the data is never at risk of leaving defined security and privacy boundaries. Data is secured with encryption, and by setting up access control, restrictions, and auditability through integration with the company’s authentication system.

Built strategically for data protection, the technology company uses LayerNext to curate and process datasets while adhering to regulations.

Data pre-processing and management

Categorizing and searching for encrypted data from several different sources can be time-consuming when done manually. In the context of machine learning, it would be impossible to sort through every permutation of surgical room, procedure, uniforms, and instrument vendors. LayerNext’s MetaLake sorts data types automatically, and features extensive metadata functionalities at every stage of the data pipeline to empower encrypted data categorization and processing.

Once the data is cleaned, and selected, automated workflows send encrypted data to the annotation team for object labeling on LayerNext’s platform. The data engineering team utilizes LayerNext to monitor annotated data, ensuring accurate representations are fed into the machine learning process.

Model accuracy

AI that operates in a healthcare context requires high accuracy. Non-uniform layouts in operating theaters, and differences in lighting, equipment, and object placements are many ways that make training a precise AI model tricky. Datasets had to be added to and balanced out so that the model can perform optimally in every procedure and environment.

LayerNext uses data and results from training and production environments to benchmark the performance of machine learning models over time, allowing data engineering teams to discover improvement opportunities on the go, ensuring optimal model performance at all times.

Through the 3 solutions above, LayerNext provides a solid foundation for a health tech company’s computer vision efforts. Ensuring HIPAA compliance while providing an improvement to data curation, management, and AI-training workflow capabilities meant the data engineering team had increased capabilities in scenario modeling and model training. This led to the development of the AI model at 4x the efficiency rate with 95% model accuracy, delivering a viable AI solution for healthcare.